merge

theano/sandbox/cuda/cuda_ndarray.cu

@@ -1412,7 +1412,7 @@ CudaNdarray_setitem(PyObject *o, PyObject  *key, PyObject  *v)
     {
         PyErr_SetString(PyExc_RuntimeError, "CudaNdarray_setitem: syncing structure to device failed");
         Py_DECREF(rval);
-        return NULL;
+        return -1;
     }
 
     CudaNdarray *viewCopyForComparison = 

theano/sandbox/cuda/opt.py

@@ -226,7 +226,12 @@ def local_gpu_sum(node):
                 gsum=GpuSum(reduce_mask)
                 pattern=(''.join(str(i) for i in reduce_mask))
                 if hasattr(gsum, 'c_code_reduce_%s'%pattern):
-                    return [host_from_gpu(gsum(gpu_from_host(x)))]
+                    rval = host_from_gpu(gsum(gpu_from_host(x)))
+                    if rval.type == node.outputs[0].type:
+                        return [rval]
+                    else:
+                        print >> sys.stderr, "WARNING: local_gpu_sum got type wrong"
+                        return None
                 else:
 
                     # Try to make a simpler pattern based on reshaping
@@ -253,9 +258,13 @@ def local_gpu_sum(node):
                         reshaped_x = x.reshape(tensor.stack(*new_in_shp))
                         sum_reshaped_x = host_from_gpu(new_gsum(gpu_from_host(reshaped_x)))
                         unreshaped_sum = sum_reshaped_x.reshape(tensor.stack(*shape_of[node.outputs[0]]))
-                        return [unreshaped_sum]
+                        if unreshaped_sum.type == node.outputs[0].type:
+                            return [unreshaped_sum]
+                        else:
+                            print >> sys.stderr, "WARNING: local_gpu_sum got type wrong"
+                            return None
 
-                    raise Exception("GpuSum don't have implemented the pattern",pattern)
+                        raise Exception("GpuSum don't have implemented the pattern",pattern)
     return False
 
 @register_opt()

theano/tensor/basic.py

@@ -484,12 +484,12 @@ class TensorType(Type):
                 return False
             if 'int' in str(a.dtype):
                 return numpy.all(a==b)
-            elif a.shape == (): #for comparing scalars, use broadcasting.
-                # Note: according to James B, there was a reason for the
-                # following two lines, that may seem weird at first glance.
-                # If someone can figure out what it is, please say it here!
-                ones = numpy.ones(2)
-                return _allclose(ones * a, ones*b)
+            #elif a.shape == (): #for comparing scalars, use broadcasting.
+                ## Note: according to James B, there was a reason for the
+                ## following two lines, that may seem weird at first glance.
+                ## If someone can figure out what it is, please say it here!
+                #ones = numpy.ones(2)
+                #return _allclose(ones * a, ones*b) ### dtype handling is wrong here
             else:
                 cmp = _allclose(a, b)
                 if cmp:

theano/tests/unittest_tools.py

@@ -63,3 +63,15 @@ def verify_grad(op, pt, n_tests=2, rng=None, *args, **kwargs):
         rng = numpy.random
     T.verify_grad(op, pt, n_tests, rng, *args, **kwargs)
 
+#
+# This supports the following syntax:
+# 
+# try:
+#     verify_grad(...)
+# except verify_grad.E_grad, e:
+#     print e.num_grad.gf
+#     print e.analytic_grad
+#     ...
+#
+#
+verify_grad.E_grad = T.verify_grad.E_grad